Add model card
Browse filesUpload fine-tuned rerankers for BioASQ 14B
Co-authored-by: André Ribeiro <andrepedro2004@hotmail.com>
Co-authored-by: Rúben Garrido <rubengarrido@ua.pt>
README.md
ADDED
|
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
license: apache-2.0
|
| 4 |
+
tags:
|
| 5 |
+
- bioasq
|
| 6 |
+
- biomedical
|
| 7 |
+
- reranking
|
| 8 |
+
- information-retrieval
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# BioASQ Phase A Reranker Models
|
| 12 |
+
|
| 13 |
+
Fine-tuned rerankers for biomedical document retrieval, trained on BioASQ data.
|
| 14 |
+
All models are cross-encoders fine-tuned from publicly available base models.
|
| 15 |
+
|
| 16 |
+
## Loading a model
|
| 17 |
+
|
| 18 |
+
```python
|
| 19 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
| 20 |
+
|
| 21 |
+
subfolder = "nvidia-llama-nemotron-rerank-1b-v2-E2-S4-Mmulti_neg_pairwise-Linfonce-FullData"
|
| 22 |
+
tokenizer = AutoTokenizer.from_pretrained("IEETA/BioASQ-14B", subfolder=subfolder)
|
| 23 |
+
model = AutoModelForSequenceClassification.from_pretrained("IEETA/BioASQ-14B", subfolder=subfolder)
|
| 24 |
+
```
|
| 25 |
+
|
| 26 |
+
For nvidia/llama-nemotron variants, also copy `llama_bidirectional_model.py` from the subfolder
|
| 27 |
+
and pass `trust_remote_code=True`.
|
| 28 |
+
|
| 29 |
+
---
|
| 30 |
+
|
| 31 |
+
## outputs-E5-Pairwise — Shifter sampler, 5 epochs, pairwise
|
| 32 |
+
|
| 33 |
+
| Model | Path | map-bioasq@10 |
|
| 34 |
+
|---|---|---|
|
| 35 |
+
| nvidia/llama-nemotron-rerank-1b-v2 | `nvidia-llama-nemotron-rerank-1b-v2-E5-Pairwise` | 0.9970 |
|
| 36 |
+
| BAAI/bge-reranker-v2-m3 | `BAAI-bge-reranker-v2-m3-E5-Pairwise` | 0.6824 |
|
| 37 |
+
| BAAI/bge-reranker-base | `BAAI-bge-reranker-base-E5-Pairwise` | 0.6686 |
|
| 38 |
+
| nboost/pt-biobert-base-msmarco | `nboost-pt-biobert-base-msmarco-E5-Pairwise` | 0.6608 |
|
| 39 |
+
| cross-encoder/ms-marco-MiniLM-L-6-v2 | `cross-encoder-ms-marco-MiniLM-L-6-v2-E5-Pairwise` | 0.6373 |
|
| 40 |
+
| ncbi/MedCPT-Cross-Encoder | `ncbi-MedCPT-Cross-Encoder-E5-Pairwise` | 0.6404 |
|
| 41 |
+
| michiyasunaga/BioLinkBERT-base | `michiyasunaga-BioLinkBERT-base-E5-Pairwise` | 0.6403 |
|
| 42 |
+
| monologg/biobert_v1.1_pubmed | `monologg-biobert_v1.1_pubmed-E5-Pairwise` | 0.6346 |
|
| 43 |
+
| microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext | `microsoft-BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext-E5-Pairwise` | 0.6291 |
|
| 44 |
+
| pritamdeka/S-PubMedBert-MS-MARCO | `pritamdeka-S-PubMedBert-MS-MARCO-E5-Pairwise` | 0.5985 |
|
| 45 |
+
| allenai/specter2_base | `allenai-specter2_base-E5-Pairwise` | 0.5912 |
|
| 46 |
+
| dmis-lab/biobert-base-cased-v1.2 | `dmis-lab-biobert-base-cased-v1.2-E5-Pairwise` | 0.5848 |
|
| 47 |
+
| cross-encoder/ms-marco-electra-base | `cross-encoder-ms-marco-electra-base-E5-Pairwise` | 0.5654 |
|
| 48 |
+
| emilyalsentzer/Bio_ClinicalBERT | `emilyalsentzer-Bio_ClinicalBERT-E5-Pairwise` | 0.4587 |
|
| 49 |
+
| cambridgeltl/SapBERT-from-PubMedBERT-fulltext | `cambridgeltl-SapBERT-from-PubMedBERT-fulltext-E5-Pairwise` | 0.2594 |
|
| 50 |
+
|
| 51 |
+
---
|
| 52 |
+
|
| 53 |
+
## outputs — Experiments
|
| 54 |
+
|
| 55 |
+
| Model | Characteristics | Path | map-bioasq@10 |
|
| 56 |
+
|---|---|---|---|
|
| 57 |
+
| nvidia/llama-nemotron-rerank-1b-v2 | E2-S4, multi_neg_pairwise, InfoNCE, FullData | `nvidia-llama-nemotron-rerank-1b-v2-E2-S4-Mmulti_neg_pairwise-Linfonce-FullData` | 0.9995 |
|
| 58 |
+
| nvidia/llama-nemotron-rerank-1b-v2 | E2, pairwise (13B1+13B2) | `nvidia-llama-nemotron-rerank-1b-v2_llama-E2-Pairwise` | 0.9970 |
|
| 59 |
+
| BAAI/bge-reranker-v2-m3 | E2-S1, pairwise, FullData, shifter | `BAAI-bge-reranker-v2-m3-E2-S1-Mpairwise-FullDataTrue` | 0.6705 |
|
| 60 |
+
| BAAI/bge-reranker-base | E2-S1, pairwise, FullData, shifter | `BAAI-bge-reranker-base-E2-S1-Mpairwise-FullDataTrue` | 0.6489 |
|
| 61 |
+
| nboost/pt-biobert-base-msmarco | E2-S1, pairwise, FullData, shifter | `nboost-pt-biobert-base-msmarco-E2-S1-Mpairwise-FullDataTrue` | 0.6274 |
|
| 62 |
+
| ncbi/MedCPT-Cross-Encoder | E2-S1, pairwise, FullData, shifter | `ncbi-MedCPT-Cross-Encoder-E2-S1-Mpairwise-FullDataTrue` | 0.6251 |
|
| 63 |
+
| michiyasunaga/BioLinkBERT-base | E2-S1, pairwise, FullData, shifter | `michiyasunaga-BioLinkBERT-base-E2-S1-Mpairwise-FullDataTrue` | 0.6178 |
|
| 64 |
+
| microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext | E2-S1, pairwise, FullData, shifter | `microsoft-BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext-E2-S1-Mpairwise-FullDataTrue` | 0.6153 |
|
| 65 |
+
| cross-encoder/ms-marco-MiniLM-L-6-v2 | E3-S8, multi_neg_pairwise | `cross-encoder-ms-marco-MiniLM-L-6-v2-E3-S8-Mmulti_neg_pairwise` | 0.6098 |
|
| 66 |
+
| monologg/biobert_v1.1_pubmed | E2-S1, pairwise, FullData, shifter | `monologg-biobert_v1.1_pubmed-E2-S1-Mpairwise-FullDataTrue` | 0.6053 |
|
| 67 |
+
| cross-encoder/ms-marco-MiniLM-L-6-v2 | E2-S1, pairwise, FullData, shifter | `cross-encoder-ms-marco-MiniLM-L-6-v2-E2-S1-Mpairwise-FullDataTrue` | 0.5944 |
|
| 68 |
+
| pritamdeka/S-PubMedBert-MS-MARCO | E2-S1, pairwise, FullData, shifter | `pritamdeka-S-PubMedBert-MS-MARCO-E2-S1-Mpairwise-FullDataTrue` | 0.5839 |
|
| 69 |
+
| michiyasunaga/BioLinkBERT-large | E2-S1, pairwise, FullData, shifter | `michiyasunaga-BioLinkBERT-large-E2-S1-Mpairwise-FullDataTrue` | 0.5781 |
|
| 70 |
+
| ncbi/MedCPT-Cross-Encoder | E3-S1, pairwise, FullData, shifter | `ncbi-MedCPT-Cross-Encoder-E3-S1-Mpairwise-FullDataTrue` | 0.5766 |
|